The PDF Is Not the Context

A PDF preserves where marks sit on a page. That is useful when the destination is a screen or printer. An agent needs something else: headings, paragraphs, tables, reading order, relationships, permissions, and a reliable way to return to the evidence. The file may contain all of that visually while exposing almost none of it cleanly.

So the model spends work reconstructing the document before it can answer the question. Columns arrive out of order. Headers repeat. Tables flatten into ambiguous strings. Scanned pages need another interpretation step. The same cleanup happens again when the next agent opens the same file, and again when a verification route checks the answer.

That is not an intelligence problem. It is an input-design failure. Buying a larger context window only gives the failure more room to run.

A software factory should ingest a document once and produce durable, governed context. Extract the text and structural units. Preserve page and region references. Identify tables without pretending uncertain cells are certain. Record which parser, vision route, and transformations touched the artifact. Then store the result behind a stable interface every authorized worker can use.

The original remains the source artifact. The normalized representation becomes production inventory. Agents retrieve the smallest relevant units instead of repeatedly receiving the whole presentation file. When a question requires layout or visual evidence, the route can return to the exact page region rather than guessing from a lossy text dump.

This is what factories are for: turn an expensive, inconsistent activity into retained machinery. The first extraction can cost more if it prevents every later worker from purchasing the same understanding again.

Teams often respond to context cost by summarizing everything. That lowers consumption, but it can also erase the qualification, footnote, exception, or table row that determines whether an answer is safe. A short summary with no route back to its evidence is cheap uncertainty wearing a clean interface.

Make every derived unit traceable to the source region that produced it. Keep extraction confidence and unresolved ambiguity in the record. Require consequential answers to cite the retained evidence internally, then make an independent stage confirm that the cited region supports the claim. If the structure cannot be recovered reliably, route the decision to a person instead of letting fluent prose conceal the gap.

The factory should be allowed to say that a document is unreadable. It should never be allowed to silently convert unreadable into true.

Token spend becomes actionable when it is attached to a mechanism. Measure ingestion, retrieval, model calls, retries, verification, and human intervention by document class and task. Keep failed extractions in the economics. A route that looks cheap because a person repairs every broken table is moving cost off the model bill and onto the most limited part of the system.

Then set policy. Reuse normalized context when the source is unchanged. Reprocess only affected units when it changes. Send visually complex pages through a stronger route and ordinary text through a lean one. Bound retrieval before the agent runs. Expire derived context when permissions or source versions change.

The goal is not to minimize tokens in isolation. The goal is to stop paying repeatedly for work the factory should remember and to know when a cheaper representation stopped being faithful.

Our position is that agent costs will force companies to confront how casually they package knowledge. Files designed for human presentation will keep entering the business. They do not have to enter every model call unchanged.

Build an ingestion factory. Preserve the original, normalize the structure, retain provenance, enforce access at retrieval, and verify claims against evidence outside the generating agent. Treat the resulting context as an asset with versions and owners, not as a pile of text copied into a prompt.

Models will get cheaper and context windows will grow. Neither improvement makes repeated decoding intelligent. The PDF is an artifact. Context is machinery. If you keep confusing them, the bill is only the most visible consequence.